lkml.org 
[lkml]   [2017]   [Apr]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] thermal: core: Add a back up thermal shutdown mechanism
Hey,

On Wed, Apr 12, 2017 at 11:31:18AM -0500, Grygorii Strashko wrote:
>
>
> On 04/12/2017 10:44 AM, Eduardo Valentin wrote:
> > Hello,
> >
> ...
>
> >
> > I agree. But there it nothing that says it is not reenterable. If you
> > saw something in this line, can you please share?
> >
> >>>> will you generate a patch to do this?
> >>> Sure. I will generate a patch to take care of 1) To make sure that
> >>> orderly_poweroff is called only once right away. I have already
> >>> tested.
> >>>
> >>> for 2) Cancel all the scheduled work queues to monitor the
> >>> temperature.
> >>> I will take some more time to make it and test.
> >>>
> >>> Is that okay? Or you want me to send both together?
> >>>
> >> I think you can send patch for step 1 first.
> >
> > I am happy to see that Keerthy found the problem with his setup and a
> > possible solution. But I have a few concerns here.
> >
> > 1. If regular shutdown process takes 10seconds, that is a ballpark that
> > thermal should never wait. orderly_poweroff() calls run_cmd() with wait
> > flag set. That means, if regular userland shutdown takes 10s, we are
> > waiting for it. Obviously this not acceptable. Specially if you setup
> > critical trip to be 125C. Now, if you properly size the critical trip to
> > fire before hotspot really reach 125C, for 10s (or the time it takes to
> > shutdown), then fine. But based on what was described in this thread,
> > his system is waiting 10s on regular shutdown, and his silicon is on
> > out-of-spec temperature for 10s, which is wrong.
> >
> > 2. The above scenario is not acceptable in a long run, specially from a
> > reliability perspective. If orderly_poweroff() has a possibility to
> > simply never return (or take too long), I would say the thermal
> > subsystem is using the wrong API.
> >
>
>
> Hh, I do not see that orderly_poweroff() will wait for anything now:
> void orderly_poweroff(bool force)
> {
> if (force) /* do not override the pending "true" */
> poweroff_force = true;
> schedule_work(&poweroff_work);
> ^^^^^^^ async call. even here can be pretty big delay if system is under pressure
> }
>
>
> static int __orderly_poweroff(bool force)
> {
> int ret;
>
> ret = run_cmd(poweroff_cmd);
> ^^^^ no wait for the process - only for exec. flags == UMH_WAIT_EXEC

Yeah, and that is what I really meant. Sorry for the confusion. The exec
is problematic in his scenario too, given he is running on a very
interesting NFS setup. Yes, the WAIT_EXEC is set:
392 static int run_cmd(const char *cmd)
393 {
394 char **argv;
395 static char *envp[] = {
396 "HOME=/",
397 "PATH=/sbin:/bin:/usr/sbin:/usr/bin",
398 NULL
399 };
400 int ret;
401 argv = argv_split(GFP_KERNEL, cmd, NULL);
402 if (argv) {
403 ret = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
404 argv_free(argv);
405 } else {
406 ret = -ENOMEM;
407 }
408
409 return ret;
410 }
411


>
> if (ret && force) {
> pr_warn("Failed to start orderly shutdown: forcing the issue\n");
>
> /*
> * I guess this should try to kick off some daemon to sync and
> * poweroff asap. Or not even bother syncing if we're doing an
> * emergency shutdown?
> */
> emergency_sync();
> kernel_power_off();
> ^^^ force power off, but only if run_cmd() failed - for example /sbin/poweroff doesn't exist
> }
>
> return ret;
> }
>
> static bool poweroff_force;
>
> static void poweroff_work_func(struct work_struct *work)
> {
> __orderly_poweroff(poweroff_force);
> }
>
> As result thermal has no control of power off any more after calling orderly_poweroff() and can get the result
> of US poweroff binary execution.
>
> >
> > If you are going to implement the above two patches, keep in mind:
> > i. At least within the thermal subsystem, you need to take care of all
> > zones that could trigger a shutdown.
> > ii. serializing the calls to orderly_poweroff() seams to be more
> > concerning than cancelling all monitoring.
> >
> >
>
> --
> regards,
> -grygorii
[unhandled content-type:application/pgp-signature]
\
 
 \ /
  Last update: 2017-04-12 18:35    [W:0.058 / U:6.356 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site